GPU Cluster & KI-Skalierung mit Kubernetes: Leitfaden 2026 für deutsche Unternehmen

# gpu-cluster-config.yaml
apiVersion: eksctl.io/v1alpha5
kind: ClusterConfig

metadata:
  name: ai-gpu-cluster
  region: eu-central-1

nodeGroups:
  - name: gpu-nodes
    instanceType: p3.2xlarge # 1x V100 GPU
    minSize: 2
    maxSize: 10
    volumeSize: 100
    labels:
      gpu: 'v100'
      ai-workload: 'inference'
    taints:
      - key: nvidia.com/gpu
        value: present
        effect: NoSchedule

  - name: gpu-training
    instanceType: p3.8xlarge # 4x V100 GPUs
    minSize: 1
    maxSize: 5
    labels:
      gpu: 'v100'
      ai-workload: 'training'

# inference-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: llm-inference
spec:
  replicas: 3
  selector:
    matchLabels:
      app: llm-inference
  template:
    metadata:
      labels:
        app: llm-inference
    spec:
      containers:
        - name: llm-server
          image: nvcr.io/nvidia/tritonserver:23.12-py3
          resources:
            limits:
              nvidia.com/gpu: 1
              memory: '16Gi'
              cpu: '8'
            requests:
              nvidia.com/gpu: 1
              memory: '12Gi'
              cpu: '4'
          ports:
            - containerPort: 8000
          env:
            - name: MODEL_REPOSITORY
              value: '/models'
          volumeMounts:
            - name: model-storage
              mountPath: /models
      volumes:
        - name: model-storage
          persistentVolumeClaim:
            claimName: model-pvc
      tolerations:
        - key: nvidia.com/gpu
          operator: Exists
          effect: NoSchedule

# gpu-hpa.yaml
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: llm-inference-hpa
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: llm-inference
  minReplicas: 2
  maxReplicas: 20
  metrics:
    - type: Resource
      resource:
        name: cpu
        target:
          type: Utilization
          averageUtilization: 70
    - type: Resource
      resource:
        name: memory
        target:
          type: Utilization
          averageUtilization: 80
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 300
      policies:
        - type: Percent
          value: 50
          periodSeconds: 60

# gpu-monitoring.yaml
apiVersion: v1
kind: ConfigMap
metadata:
  name: gpu-prometheus-config
data:
  prometheus.yml: |
    global:
      scrape_interval: 15s
    scrape_configs:
    - job_name: 'gpu-metrics'
      static_configs:
      - targets: ['gpu-exporter:9445']
      metrics_path: /metrics
    - job_name: 'kubernetes-pods'
      kubernetes_sd_configs:
      - role: pod
      relabel_configs:
      - source_labels: [__meta_kubernetes_pod_annotation_prometheus_io_scrape]
        action: keep
        regex: true

# workload-scheduler.py
import kubernetes
from typing import Dict, List

class GPUScheduler:
    def __init__(self):
        self.on_prem_gpus = 8  # A100 Cluster
        self.cloud_gpus = 20   # Spot Instances

    def schedule_workload(self, workload: Dict) -> str:
        if workload['priority'] == 'high':
            return 'on-prem'  # Garantierte Verfügbarkeit

        if workload['batch_size'] > 1000:
            return 'cloud-spot'  # Kostengünstig für große Batches

        if self.on_prem_utilization < 0.7:
            return 'on-prem'  # Lokale Kapazität nutzen

        return 'cloud-on-demand'  # Fallback

# multi-cluster-config.yaml
apiVersion: v1
kind: ConfigMap
metadata:
  name: cluster-config
data:
  clusters.yaml: |
    clusters:
      - name: on-prem-gpu
        endpoint: https://on-prem-k8s:6443
        gpu-count: 8
        gpu-type: "a100"
        cost-per-hour: 0.40
        
      - name: aws-gpu-spot
        endpoint: https://aws-eks:6443
        gpu-count: 20
        gpu-type: "a100"
        cost-per-hour: 0.75
        
      - name: azure-gpu-on-demand
        endpoint: https://azure-aks:6443
        gpu-count: 10
        gpu-type: "a100"
        cost-per-hour: 2.30